Controlling Torque Versus Position: A Direct Comparison of Ankle Exoskeleton Assistance Strategies

Abstract

Lower-limb exoskeletons are typically controlled using either torque control, where per-stride joint torques are specified and kinematics emerge through user-device interaction, or position control, in which joint kinematics are specified and torques emerge. Here, we compare these strategies in an ankle exoskeleton with ten experienced exoskeleton users. Participants first walked using a torque control profile previously shown to substantially reduce metabolic rate. Using the measured stance phase kinematics from torque control, participants next walked with position control, which enforced the same kinematics as a function of percent stance. We found that, despite matched kinematics, kinetics differed: peak exoskeleton torque was 50% greater and occurred 10% later in stance with torque control than with position control. The metabolic rate reduction relative to walking in the device unpowered was 22% for torque control and 8% for position control. Step time variability was 27% larger with torque control versus position control. Most participants (7 of 10) preferred the position controller, reporting differences in perceived comfort but not effort. Even with matched kinematics, position and torque control interact with users in ways that result in significant differences in kinetics, energy economy, balance, and comfort, and both controllers offer potential benefits that merit further study.

Methods

Torque Controller

\[ \dot{\theta}^{motor}_{des} = k_P \cdot e + k_D \cdot \dot{e} + u_{learn}(i_{stride} + D) \]

Exoskeleton torque was prescribed as a phase-based profile: proportional term $k_P \cdot e$ corrected current error, derivative term $k_D \cdot\dot{e}$ reacted to error changes, and feedforward learning term $u_{learn}(i_{stride} + D)$ improve tracking using past errors. Learning term was offset by delay $D$ to account for transmission latency, and was updated each step as $u_{learn}(i_{stance}) \xleftarrow{} u_{learn}(i_{stance}) \cdot \beta + k_L \cdot e$, with forgetting factor $\beta = 0.99$. Gains $k_P$, $k_D$, and $k_L$ were set empirically.

Position Controller

\[ \theta^{motor}_{des} = f_{nominal}(\theta^{ankle}_{des}) + k_P \cdot e + \alpha \cdot u_{learn}(i_{stride} + D) \]

Ankle angle was prescribed as a phase-varying trajectory: nominal mapping $f_{nominal}(\theta^{ankle}_{des})$ empirically mapped desired ankle angle to motor position, proportional term $k_P \cdot e$ corrected current error, and scaled learning term $\alpha \cdot u_{learn}(i_{stride} + D)$ improved tracking as in the torque controller, with scalar $\alpha$ ramping from 0 to 1 over the first 20% of stance to limit instability at heel strike and applied torque before foot-flat.

Results

Kinematics / Kinetics

Despite matched kinematics, peak torque applied when walking in torque control is 50.3% higher than when walking in position control, and occurs 10.5% later in stance. Torque control exhibited greater range of motion, torque magnitude, and post-toe-off joint velocity.

Metabolic Rate Outcomes

Participants' survey responses indicated a preference for position control despite smaller metabolic rate reductions. suggesting that factors beyond energy economy contribute to user preference. One likely contributor is stability: position control resulted in lower variability, which may have been perceived as less destabilizing. Predictability may also play a role, as the position controller delivers torque as a function of joint position, which may be more intuitive than torque control delivering torque based on time since heel strike.

User Preference / Variability

On average, the metabolic rate reduction with position control was approximately one-third that with torque control, and torque control yielded greater reductions in nearly all individuals. Participants who experienced larger metabolic rate reductions with torque control also had larger reductions with position control. Participants walking with position control did not modify their gait to generate larger exoskeleton torques even though doing so appeared possible with this controller, and such larger torques would likely have led to a larger metabolic rate reduction.

BibTeX

@article{martin2026tqvspos,
  author    = {Martin, Russell M. and Punamiya, Rohan S. and Collins, Steven H.},
  title     = {Controlling Torque Versus Position: A Direct Comparison of Ankle Exoskeleton Assistance Strategies},
  journal   = {TMRB},
  year      = {2026},
}